v2: Fix pasteo of an extra abs being inserted (caught by many). Rewrite
to drop the silly switch statement.
Reviewed-by: Matt Turner <mattst88@gmail.com> (v1)
On Haswell, POW takes 24 cycles, while EXP2 only takes 14. Plus, using
POW requires putting 2.0 in a register, while EXP2 doesn't.
I believe that EXP2 will be faster than POW on basically all GPUs, so
it makes sense to optimize it.
Looking at the savage2 subset of shader-db:
total instructions in shared programs: 113225 -> 113179 (-0.04%)
instructions in affected programs: 2139 -> 2093 (-2.15%)
instances of 'math pow': 795 -> 749 (-6.14%)
instances of 'math exp': 389 -> 435 (11.8%)
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Surprisingly, this helps one vertex shader in 3DMMES.
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The comment was stale, because the lowering in question wasn't happening
in lower_instructions.cpp. Presumably if the lowering ever moves there,
we can plumb the lowering mask through to opt_algebraic.
total instructions in shared programs: 1618696 -> 1616810 (-0.12%)
instructions in affected programs: 243018 -> 241132 (-0.78%)
GAINED: 0
LOST: 0
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
I noticed this in a shader in Unigine Heaven that was spilling. While it
doesn't really reduce register pressure, it shaves a few instructions
anyway (7955 -> 7882).
v2: Fix turning "0 >> x" into "x" instead of "0" (caught by Erik
Faye-Lund).
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
While ir_builder is slightly less efficient, we're only increasing the
work when there's actual optimization being done, and it's way more
readable code.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Matt and I had each screwed up these common required patterns recently, in
ways that wouldn't have been noticed for a long time if not for code
review. Just enforce it in the caller so that we don't rely on code
review catching these bugs.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Two extra instructions in some heroesofnewerth shaders, but a win for
everything else.
total instructions in shared programs: 1531352 -> 1530815 (-0.04%)
instructions in affected programs: 121898 -> 121361 (-0.44%)
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Many GPUs have an instruction to do linear interpolation which is more
efficient than simply performing the algebra necessary (two multiplies,
an add, and a subtract).
Pattern matching or peepholing this is more desirable, but can be
tricky. By using an opcode, we can at least make shaders which use the
mix() built-in get the more efficient behavior.
Currently, all consumers lower ir_triop_lrp. Subsequent patches will
actually generate different code.
v2 [mattst88]:
- Add LRP_TO_ARITH flag to ir_to_mesa.cpp. Will be removed in a
subsequent patch and ir_triop_lrp translated directly.
v3 [mattst88]:
- Move changes from the next patch to opt_algebraic.cpp to accept
3-src operations.
Reviewed-by: Matt Turner <mattst88@gmail.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Because these classes are used entirely from their own source files
and not from separate DSOs, the linker gets to produce massively less
code. This cuts about 13k of text in the libdricore case. In the
non-libdricore case, the additional linkage information allows the
compiler to inline some code, so libglsl.a size actually increases by
about 300 bytes.
For a dricore build, improves shader_runner runtime on
glsl-fs-copy-propagation-texcoords-1 by 0.21% +/- 0.03% (n=353574,
outliers removed). No statistically significant difference with n=322
on glslparsertest on a yofrankie shader intended to test compiler
performance.
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
The vector operator collects 2, 3, or 4 scalar components into a
vector. Doing this has several advantages. First, it will make
ud-chain tracking for components of vectors much easier. Second, a
later optimization pass could collect scalars into vectors to allow
generation of SWZ instructions (or similar as operands to other
instructions on R200 and i915). It also enables an easy way to
generate IR for SWZ instructions in the ARB_vertex_program assembler.
